Laboratory of Information Technologies

Date and Time: Wednesday, 7 June 2023, at 3:00 PM

Venue: room 310, Laboratory of Information Technologies, online on Webinar

Information about seminar and link to join online

Seminar topic: “Application of the Hopfield network to SPD track reconstruction”

Speaker: Ivan Kadochnikov

Authors: Ivan Kadochnikov, Alexander Kovalenko, Gennady Ososkov

Abstract:

One of the key steps of processing experiment data in particle physics is the reconstruction of particle trajectories (tracks). For each particle interaction event the aim is to figure out which hits (points in space where some particle was detected) were generated by the same particle. In the upcoming SPD experiment on NICA particular complications will arise from the high frequency of events (3 MHz) and a strong background of fake hits caused by the specifics of the tracking detectors being used.As far back as the 80s Denby and Peterson proposed using a Hopfiled network (a fully connected recurrent network of binary neurons with as symmetric weight matrix) for tracking. A neuron in their network corresponds to a pair of hits, that is, a segment of a possible track. The weights of the network are set to ensure a smooth trajectory, that is a small angle between adjacent segments, rewarding smooth and non-branching tracks. While neuron activations move towards an equilibrium point, the energy function of the Hopfield network approaches a minimum, which correspond to a good tracking result. Using simulated annealing allows to search for a global network energy minimum. Despite the initial success, the interest in using Hopfield for tracking faded quickly due to slow convergence, high sensitivity to background noise, and most importantly event multiplicity caused by the rapid increase of particle beam luminosity and advances of experimental technologies. However, recently, thanks to the progress of quantum computers and quantum annealing methods, which can drastically accelerate the evolution of Hopfield networks, the interest in Hopfield neural network tracking has been renewed. In this work, we investigated the applicability of Hopfield network tracking to a set of simulated events, representing those expected in the SPD experiment. To achieve a good result, it was necessary to fine-tune weight parameters and the energy function of the network using metaparameter optimization software.